Center for Information Services and High Performance Computing Edge Bundling for Visualizing Communication Behavior Ronny Brendel, Michael Heyde, Holger Brunst, Tobias Hilbrich and Matthias Weber Presenter: Jens Domke VPA Workshop, Salt Lake City, 18 th November, 2016
Contents • Introduction & Motivation • Test Cases • Edge Bundling • Time-based Visualization • Summary Visualization • Conclusion & Future Work • References Slides and supplemental material: https://github.com/hydroo/ sc16-vpa-edge-bundling-for-visualizing-communication-behavior 2
Contents • Introduction & Motivation • Test Cases • Edge Bundling • Time-based Visualization • Summary Visualization • Conclusion & Future Work • References 3
Introduction & Motivation • Utilize today‘s computers increasing concurrency → • HPC systems use distributed memory inter-process → communication (IPC) required (e.g. Message Passing) • Analysis and visualization of IPC supports developers in their debugging and performance optimization workflows • Common tasks include detecting: • Incorrect communication, • Slow messages, • Load imbalances, and • Inefficient communication patterns 4
Introduction & Motivation • Timeline visualizations [Lamport,1978] (Paraver [Pillet,1995] , Intel Trace Analyzer [ITA,2016] , Vampir [Brunst,2014] ) • Helpful message visualization at small scales • But unfavorable visualization for many ranks 5
Introduction & Motivation • Current solution in Vampir: Message Burst • Start & end with the number of exchanged messages (varying circle size) encoded • Sender-oriented (target and duration are missing) 6
Introduction & Motivation • No obvious patterns • Goal: • Alleviate message clutter and information occlusion • Preserve interesting information • Highlight & amplify communication patterns • Approach based on ideas from edge bundling [Holten,2006] 7
Contents • Introduction & Motivation • Test Cases • Edge Bundling • Time-based Visualization • Summary Visualization • Conclusion & Future Work • References 8
Test Cases • Weather Research & Forecasting (WRF) [Michalakes,2001] • 16 processes • 4 minutes • 160k point-to-point messages • https://www.vampir.eu/public/files/tracefiles/Large.zip 9
Test Cases • FD4 [Lieber,2010] • 1,024 processes, 5 minutes, 2.2M messages 10
Test Cases • Tachyon [SPEC,2016] • 256 processes, 20 minutes, 8,000 messages 11
Contents • Introduction & Motivation • Test Cases • Edge Bundling • Time-based Visualization • Summary Visualization • Conclusion & Future Work • References 12
Edge Bundling > Time-based Vis. • Hierarchical edge bundling [Holten,2006] • B-Splines • Send/receive event points + mean shift clustering = hierarchy • Sender, Receiver (gradient) • WRF 13
Edge Bundling > Time-based Vis. + Occupies less space – Extreme Bending towards the center of the diagram 14
Edge Bundling > Time-based Vis. • Grid-based Routing [Lambert,2010] + diagonals: WRF, Tachyon 15
Edge Bundling > Time-based Vis. + Less horizontal bending + Cleaned up representation + More intuitive – Fixed position and size grid – Still some horizontal bending 16
Edge Bundling > Time-based Vis. • Parameter tuning is hard (line width, grid/hierarchy parameters, etc.) • Highly individual for each trace • But important for diagram quality • Can be very slow • Invented for static image vis. • FD4 trace takes >20 min ➡ Try edge bundling for msg. summery 17
Contents • Introduction & Motivation • Test Cases • Edge Bundling • Time-based Visualization • Summary Visualization • Conclusion & Future Work • References 18
Edge Bundling > Summary Vis. • Existing: Communication Matrix: WRF, FD4, Tachyon • Encode quantity into the points in color (here number of messages exchanged) 19
Edge Bundling > Summary Vis. • Alternatively, use chord diagrams • Sender, Receiver • Additivate color mixing: Not always helpful, but sometimes WRF FD4 Tachyon 20
Edge Bundling > Summary Vis. • Sender/receiver diagram • Sending process left, receiver right • Grid with diagonals • YX-Routing preferring diagonals • One color for each sender • WRF: Near communication only • FD4: Mostly near communication, some farther away receivers 21
Edge Bundling > Summary Vis. • Tachyon: all-to-one • LULESH [Karlin,2013] : all-to-all 22
Edge Bundling > Summary Vis. Advantages • Some comm. patterns easily identifyable • Good alternative to matrix view or chord diagrams Disadvanages • Parameter choice is challenging for also for summery vis. • Implementation still rel. slow (FD4 appr. 30s) 23
Contents • Introduction & Motivation • Test Cases • Edge Bundling • Time-based Visualization • Summary Visualization • Conclusion & Future Work • References 24
Conclusion • Presented edge bundling for time-based and summary inter- process communication visualization • Time-based Visualization: • Hierarchical produces unintuitive results • Grid-based approach better • Introduced the sender/receiver diagram: • Simpler and faster • Improved ability to highlight communication patterns over other summary based visualizations • Edge bundling proved challenging (parameter choice, tuning, ...) 25
Future Work • Improve visual outcome of the sender/receiver diagram • Encode quantities into edges • Alternative ways to obtain control polygons • Make edge bundling viable for time-based visualizations • Reduce message number via filtering • Flexible grid to improve image quality • Parallel implementation for runtime improvement • Explore alternative edge bundling techniques, e.g. force-directed edge bundling [Holten,2009] Slides and supplemental material: https://github.com/hydroo/ sc16-vpa-edge-bundling-for-visualizing-communication-behavior 26
References • [Lamport,1978] L. Lamport, “Time Clocks, and the Ordering of Events in a Distributed System,” Commununications of the ACM, vol. 21, pp. 558–565, 1978. • [Pillet,1995] V. Pillet, J. Labarta, T. Cortes, and S. Girona, “Paraver: A tool to visualize and analyze parallel code,” in Proceedings of WoTUG 18: Transputer and occam Developments, vol. 44, 1995, pp. 17–31. • [ITA,2016] “Intel Trace Analyzer and Collector,” https://software.intel.com/intel-trace-analyzer, Jul. 2016. • [Brunst,2014] H. Brunst and M. Weber, “Custom hot spot analysis of HPC software with the Vampir performance tool suite,” in Tools for High Performance Computing 2012. Springer, 2013, pp. 95–114. • [Holten,2006] D. Holten, “Hierarchical edge bundles: Visualization of adjacency relations in hierarchical data,” IEEE Transactions on Visualization and Computer Graphics, vol. 12, no. 5, pp. 741 748, 2006. • [Michalakes,2001] J. Michalakes, S. Chen, J. Dudhia, L. Hart, J. Klemp, J. Middlecoff, and W. Skamarock, “Development of a next generation regional weather research and forecast model,” in Developments in Teracomputing: Proceedings of the Ninth ECMWF Workshop on the use of high performance 27 computing in meteorology, vol. 1. World Scientific, 2001, pp. 269–276.
References • [Lieber,2010] M. Lieber, V. Grủtzun, R. Wolke, M. S. Mủller, and W. E. Nagel, “Highly scalable dynamic load balancing in the atmospheric modeling system COSMO-SPECS+FD4,” in International Workshop on Applied Parallel Computing. Springer, 2010, pp. 131–141. • [SPEC,2016] “122.tachyon SPEC MPI2007 benchmark description,” http://www.spec.org/mpi2007/docs/122.tachyon.html, Jun. 2016. • [Lambert,2010] A. Lambert, R. Bourqui, and D. Auber, “Winding roads: Routing edges into bundles,” in Computer Graphics Forum, vol. 29, no. 3. Wiley Online Library, 2010, pp. 853–862. • [Karlin,2013] I. Karlin, A. Bhatele, J. Keasler, B. L. Chamberlain, J. Cohen, Z. DeVito, R. Haque, D. Laney, E. Luke, F. Wang et al., “Exploring traditional and emerging parallel programming models using a proxy application,” in Parallel & Distributed Processing (IPDPS), 2013 IEEE 27th International Symposium on. IEEE, 2013, pp. 919–932. • [Holten,2009] D. Holten and J. J. Van Wijk, “Force-directed edge bundling for graph visualization,” in Computer Graphics Forum, vol. 28, no. 3. Wiley Online Library, 2009, pp. 983–990. 28
Tachyon with Hierarchical Edge Bundl. 29
Recommend
More recommend
